| Author: | Wang, Yuxuan |
| Title: | Study of artificial intelligence for visual defect inspection in industrial products |
| Advisors: | Chung, S. H. Nick (ISE) |
| Degree: | Ph.D. |
| Year: | 2025 |
| Subject: | Engineering inspection -- Automation Computer vision -- Industrial applications Quality control -- Optical methods Hong Kong Polytechnic University -- Dissertations |
| Department: | Department of Industrial and Systems Engineering |
| Pages: | xii, 111 pages : color illustrations |
| Language: | English |
| Abstract: | Visual defect inspection is a critical and challenging task in the quality control of industrial manufacturing, essential for production efficiency, economic benefit, and end-user safety. Amidst the rapid advancements in industrial cyber-physical systems and computer vision techniques, deep learning has demonstrated significant competence in feature extraction and pattern recognition, showing great potential for classifying and locating defects in industrial products. However, challenges such as insufficient data records and diverse defect types restrict the detection accuracy of data-driven deep convolutional neural networks (CNNs). Most defect inspection studies in the literature focus on modifying existing CNN structures, operating under the assumption that ample defective data can be obtained and utilized for training inspection models. This assumption, however, does not adequately reflect the realistic conditions of many industrial products. To advance modern deep learning-based visual inspection, this thesis identifies three industrial scenarios from the perspective of data acquisition: 1) relatively sufficient defective products, 2) limited defective products, and 3) sufficient non-defective products. This thesis initially introduces a novel three-step automatic data augmentation framework designed to address the first scenario of defect inspection. The proposed framework simultaneously optimizes data augmentation policies and the neural network for industrial defect detection. A light search space is initially formulated to efficiently sample augmentation policies and generate augmented images for joint optimization. To minimize the hyperparameter tuning efforts typically associated with retraining using the identified policies, a three-step bi-level optimization scheme is proposed. This scheme replaces the traditional retraining strategy by alternately updating the model and augmentation parameters. To address the non-differentiable nature of the joint optimization scheme, policy gradient sampling is employed to efficiently estimate the gradient flow. Experimental results on three industrial defect detection datasets demonstrate that the proposed automatic augmentation framework surpasses state-of-the-art augmentation methods and significantly enhances the accuracy of the baseline defect detection model. Furthermore, the framework effectively mitigates missed defect detection in four practical industrial scenarios: textured backgrounds, uneven brightness, low contrast, and intraclass differences. Secondly, this thesis presents a few-shot defect detection case study for medical catheter defect detection to address the second defect inspection scenario. Initially, a fine-tuning-based few-shot learning scheme is introduced to gain knowledge from an abundant base dataset. Subsequently, an enlarged scale feature pyramid network (ESFPN) is designed to cover the variant sizes of defects. Finally, a contrastive proposal memory bank (CPMB) is proposed to alleviate the intraclass variation problem caused by different viewpoints and efficiently utilize similar features. Experimental results on the collected medical catheter defect dataset demonstrate the superior performance of the proposed method compared to other existing prevalent methods. Lastly, this thesis proposes a three-step unsupervised anomaly detection framework for bobbin defect inspection to address the third defect inspection scenario. Existing supervised methods encounter difficulties due to class imbalance, with defective samples being significantly less common than non-defective ones in realistic manufacturing environments. Conversely, although unsupervised methods often demonstrate robust performance, they tend to overlook inherent constraints such as the diminutive size of defects and the prevalent neglect of leveraging prior knowledge without annotations. To address these challenges, an unsupervised image anomaly detection framework for bobbin components is proposed, structured into a three-step methodology. First, the Contrastive Language-Image Pretraining (CLIP)-based attention map generation (CAMG) module creates language-based attention maps for guidance. Second, teacher-student networks extract multi-scale features, which are then integrated by the multi-scale concurrent feature integration (MCFI) module to identify anomalies across various scales. Finally, the CLIP-based attentional knowledge distillation (CAKD) module leverages the attention maps from CAMG and the multi-scale features from MCFI to enhance anomaly detection capabilities. Experimental results on the collected bobbin anomaly dataset demonstrate that the proposed framework outperforms existing methods, highlighting its superior performance in anomaly detection. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/13571